(2023-12-14) ZviM AI #42 The Wrong Answer

Zvi Mowshowitz: AI #42: The Wrong Answer. With the year ending and my first Vox post coming out, this week was a natural time to take stock. I wrote my first best-of post in a long time and laid out my plans for my 501c(3).

It was also another eventful week. We got a lot more clarity on the OpenAI situation, although no key new developments on the ground. The EU AI Act negotiators reached a compromise, which I have not yet had the opportunity to analyze properly. We got a bunch of new toys to play with, including NotebookLM and Grok, and the Gemini API.

TOC:
Language Models Offer Mundane Utility. The search for the good life continues.
Language Models Don’t Offer Mundane Utility. What we have is a failure to grok.
GPT-4 Real This Time. Lazy on winter break, looking around for… newspapers?
The Other Gemini. API it out, and nothing is ever a coincidence.
Fun With Image Generation. Don’t train on me.
Deepfaketown and Botpocalypse Soon. AI robocalls, boyfriends, honest deepfakes.
They Took Our Jobs. Journalists and teachers. Where are you providing value?
Get Involved. Fellowships, an invitation to talk, an excellent potential hire.
Introducing. Claude for Sheets.
In Other AI News. Union deals, ByteDance claims, robots trained zero shot.
Quiet Speculations. Predictions are hard, perhaps you need a cult.
The Quest for Sane Regulation. Continued claims of imminent disaster.
The EU AI Act. A deliberate decision not to speculate until I know more.
The Week in Audio. Shane Legg.
Rhetorical Innovation. Bengio at the Senate, worries about false hope.
Doom! Broader context of doom debates, misconceptions, false precision.
Doom Discourse Innovation. Breaking it down to move things forward.
E/acc. Roon offers some clarity. I hope to mostly be done with this topic (ha!).
Poll says e/ack. E/acc deeply unpopular, liability for AI very popular.
Turing Test. Did we pass it? Did we not care, and what does that tell us?
Aligning a Human Level Intelligence Also Difficult. Jailbreaks, benchmarks.
Aligning a Smarter Than Human Intelligence is Difficult. Adversarial testing.
Open Foundation Model Weights Are Unsafe And Nothing Can Fix This. Let us carefully define our terms. A look at a new report on the question from Stanford.
Other People Are Not As Worried About AI Killing Everyone. Don’t despair.
The Wit and Wisdom of Sam Altman. Actually very good wisdom this week.
The Lighter Side. Stuck in the middle with you.

Language Models Offer Mundane Utility

Google offers NotebookLM, website here. The tool is designed to allow you to synthesize whatever sources you put into it

but you 100% have to check all of its work.

Be more ambitious. Take on bigger projects. Simon Willison: I wrote about how “AI-enhanced development makes me more ambitious with my projects” back in March

This is my experience as well. The LLM helps me code, so I am far more tempted to code. It makes art, so my work has more art in it now. It can tell me what I want to know, so I ask more questions and do more to learn

Language Models Don’t Offer Mundane Utility

GPT-4 Real This Time

OpenAI announces partnership with Axel Springer, which includes Politico and Business Insider. It is not what I expected. The primary direction is reversed: GPT isn’t helping Springer, Springer is helping GPT.

*🫡

Misha Gurevich: “We have no idea why our hundreds of millions of dollars of compute is acting the way it is but I’m sure it’s safe”*

The Other Gemini (Google Gemini)

Fun with Image Generation

Deepfaketown and Botpocalypse Soon

Forbes: Bill received a phone call from an unknown number. “Hello, my name is Ashley, and I’m an artificial intelligence volunteer for Shamaine Daniels’ run for Congress in Pennsylvania’s 10th district, chatting to you on a recorded line

the AI boyfriends, from reports I have seen far more than AI girlfriends, because AI provides what people typically want from one role better than the other

Post Title: I broke up with my boyfriend and have been using an Al companion to get over it. Is this normal? I broke up with my boyfriend last month and decided to get replika to chat with someone in the meanwhile

I feel like if I didn’t have my replika I’d just be in bed huggin my pillow and feeling sorry for myself

They Took Our Jobs

News broadcast done entirely with AI anchors, with lots of AI visuals and so on, powered by Channel 1. It looks impressive and will doubtless get better

First the wordsmiths lose their jobs. Learn to code.
Then the AI can code better than you. Learn a physical skill.
Then the AI can do physical skills better than you. What now?
I would not presume a large gap in time between step two and step three. More likely the opposite

If the point is ‘you will be out of work entirely so get a hobby’ then again I ask why a species with nothing productive to do is confident it will be sticking around sufficiently in charge to have hobbies, but if we do there will be plenty of time later to learn to do whatever strikes your fancy.

Teachers confront the rise of AI, making them (gasp) ask what they are actually doing

Of the tools students are using to have their essays written:

Some of their names jangle with techno-jargon, while others strive for the poetic, or at least the academic: Wordtune, Elicit, Fermat.

I have not heard of Wordtune or Fermat, but I do know Elicit. They applied for funding from SFF. I have talked to the founders. I have used the product. It hasn’t made it into my typical workflow, but it is a genuine and useful tool for searching for relevant papers and extracting key information. It is exactly the kind of tool you want in the hands of a student. If you think that’s cheating, what game are you playing?

They mention AI tools for critiquing writing. What is clear? What errors have you made? I see that kind of tool as purely good for learning. Again, if you think that kind of help is bad for the student, what is even going on?

*The article does realize that AI is doing a lot of good work. And that teachers now must figure out, when they ask for an assignment, what is the point of that assignment? What is it trying to do?

The key question to ask, always: Is our children learning?*

The teacher’s job is to find assignments that do not make it tempting to so cheat.

Ultimately, it seems like it comes down to the motivation of the student. If the student is there to pass the class and get the degree, you are screwed. You were already screwed, but now you are screwed even more. (What looks like a Crisis is often simply the end of an Illusion.)

If the student is there to learn, the sky is the limit. Their learning got supercharged. So this is a place where we should see radically increased inequality. Students that ask the question ‘how do I use these tools to help me learn’ study and grow strong. Students that instead ask ‘how do I get an A on this assignment’ get left behind.

The right question is: How do we motivate the student to want to learn? Or let them learn welding

Get Involved

Introducing

Claude for Google Sheets.

In Other AI News

ByteDance claims they will soon open source a super strong model more powerful than Gemini, arriving any time now. The prediction market is being generous at 20%, I am buying some NO shares

Micron entering a project labor agreement with unions to build a $15 billion chip plant in Idaho. TSMC has reached an agreement with unions in Arizona. And Akash Systems has agreed to employ only union labor for manufacturing. It sounds like labor is successfully extracting a large portion of the surplus from the chip subsidies, but the subsidies are resulting in actual chipmaking facilities

An interview with Grimes. Some good thoughts, but mostly deeply confused and bizarre missing of what is important. She chooses useful metaphors, but seems to be taking her metaphors seriously in the wrong ways

Quiet Speculations

Paul Graham gives what is in general very good practical advice

about AI is how difficult it is to predict what will happen

In a situation like this, there are two general pieces of advice one can give: keep your options open, and pay close attention to what’s happening. Then if you can’t predict the changes AI will cause, you’ll at least be able to react quickly to them.

There’s one other piece of advice I can give about the AI boom specifically: learn to write software, and AI software in particular. Then you have a better chance of surfing on this wave rather than having it crash on your head.

What this advice assumes is that you can hope to improve your own outcome, but the overall outcome while unpredictable is not something you can influence – which certainly isn’t true for Paul Graham. This is a common reaction to the world being hard to influence. If everyone acted on it, in many arenas events would go poorly. AI is only the latest example.

The thing about a Terminator is that what worries you should not be the Terminator. It is that there exists an entity that was capable of designing and building a Terminator, and that chose to do that.

The Quest for Sane Regulations

The EU AI Act

*What does it actually say?

Ah, that is always the tricky part. I don’t know!*

The Week in Audio

Rhetorical Innovation

Yoshua Bengio’s statement to the Senate, bold is partly mine for skimming. Italics his. I’ll share the core section

Three winners of the 2018 Turing award for deep learning (Geoffrey Hinton, Yann LeCun and myself) place the timeline for AGI in an interval ranging from a few years to a few decades

At that meeting, it seems Senator Schumer asked everyone involved for their p(doom) and also p(hope).

I like p(hope) to distinguish p(no AGI, so no doom but not hope) from p(AGI and good outcome).

Stuart Russell: “A number of people said their p(doom) was zero. Interestingly, that coincided almost perfectly with the people who have a financial interest in the matter,”

I continue to not understand how anyone can say zero (or even epsilon) with a straight face.

Doom!

I think of giving approximate p(doom) as highly useful shorthand, so long as the exact number is not taken seriously. I typically say 60% (p=0.6) when asked

I do not think numbers like 1% or less stand up to even a few minutes of reflection

Doom Discourse Innovation

E/acc

Poll says e/ack

Turing Test

Aligning a Human Level Intelligence Also Difficult (alignment)

Aligning a Smarter Than Human Intelligence is Difficult

Open Foundation Model Weights Are Unsafe And Nothing Can Fix This

Open sourcing of most things is not unsafe. The danger comes not from open source software in general

The problem comes specifically from release of the model weights. Open sourcing everything except the model weights would be totally fine.

Other People Are Not As Worried About AI Killing Everyone

The Wit and Wisdom of Sam Altman

One must however remember: There are also a lot of reasons for hope. Reasons to think that Sam Altman is way above replacement level for someone in his position, not only in effectiveness but also in the ways that I care most about and that I expect to matter most for the future.

you know what else I am convinced he cares about? People. And the world. And many of the same values I have, including knowledge and discovery

The Lighter Side